Historically, knowledge about chemical reactions was gleaned by generalizing from relatively small sets of experimental data using the functional group concept. Often all or almost all the available data was used as a “training set” for the model, there was only a very small test set (if any). Theory was used mostly to qualitatively confirm human-proposed explanations of experimental data; it was seldom accurate enough for quantitative comparisons. New types of reactions were discovered only rarely, usually by serendipity in the laboratory when unexpected products were formed. Within this historical context, developing accurate reaction mechanisms and kinetic models, or designing efficient syntheses of new molecules, was a very challenging task even for expert chemists.
Now new ways of approaching chemical reactions are feasible. Quantum chemistry has become so accurate that it is routinely used to quantitatively determine some molecular and reaction parameters, and it can completely replace experiment in some situations. Indeed, one can now use quantum chemistry to systematically discover new types of reactions on the computer, rather than relying on lucky experimental results. Machine learning now makes it possible to effectively use almost the whole set of experimental data reported in the literature, and to have the computer rather than a human make the generalizations, not confined to conventional functional group models. With these new capabilities, in principle computer methods can be designed to handle the challenging tasks of designing efficient organic syntheses, or quantitatively predicting reaction mechanisms and kinetics. Recent progress towards establishing these new approaches to reactive chemistry will be outlined, highlighting some of the technical and conceptual challenges that are the focus of our current research.